Combinatorial Max-Flow: Algorithms & Advances

• Combinatorial max-flow is a network optimization method that computes maximum flow in discrete graphs using augmenting paths and circuit augmentation techniques.
• It unifies classical algorithms like Ford–Fulkerson, Edmonds–Karp, and push–relabel as circuit walks within a pseudoflow polyhedron, enhancing theoretical and practical insights.
• Recent advances include deterministic approaches for dense graphs, extensions to higher-dimensional complexes, and applications in computer vision and network science.

Combinatorial Max-Flow is a foundational concept in network optimization and discrete mathematics, concerning the computation of maximal commodity transfer between designated sources and sinks under explicit combinatorial constraints. Unlike continuous or numerical approaches, combinatorial max-flow focuses on algorithms and frameworks exploiting the underlying discrete network structure—typically, graphs or their higher-dimensional analogues. This area has propelled advances in both algorithmic theory and applications across scientific computation, network science, computer vision, and beyond.

1. Classical Problem Formulation and Polyhedral Structure

Consider a directed graph $G=(V,E)$ with capacity function $c:E\to\mathbb{Z}_+$ (or $\mathbb{R}_{\geq0}$), a source $s$, and a sink $t$. A flow $f:E \to \mathbb{R}_+$ respects:

• Capacity constraints: $0 \leq f(e) \leq c(e)\ \forall e \in E$
• Conservation: $$\sum_{e \in \delta^+(v)} f(e) = \sum_{e \in \delta^-(v)} f(e)$$ for all $v \in V\setminus\{s,t\}$

The goal is to maximize $$|f| = \sum_{e\in \delta^+(s)} f(e) - \sum_{e\in \delta^-(s)} f(e)$$. The max-flow/min-cut theorem asserts that this value equals the minimum capacity of an $c:E\to\mathbb{Z}_+$0-$c:E\to\mathbb{Z}_+$1 cut (Kannappan et al., 8 Nov 2025, Borgwardt et al., 2023).

This classical problem, and its solution space, naturally reside within a polyhedral framework. The feasible region (“flow polyhedron”) is a bounded, convex polyhedron, and, as shown in circuit augmentation studies, various combinatorial algorithms (augmenting path, push-relabel, etc.) can be interpreted as discrete circuit walks within this polyhedron (Borgwardt et al., 2023).

2. Canonical Combinatorial Algorithms and Circuit Augmentation

Historically, combinatorial max-flow algorithms progress by iteratively locating and pushing flow along augmenting structures defined over the residual graph:

• Generic Augmenting Path (Ford–Fulkerson): Augment along arbitrary $c:E\to\mathbb{Z}_+$2-$c:E\to\mathbb{Z}_+$3 paths until no such path remains.
• Shortest Augmenting Path (Edmonds–Karp): Restrict to shortest-length $c:E\to\mathbb{Z}_+$4-$c:E\to\mathbb{Z}_+$5 paths, yielding $c:E\to\mathbb{Z}_+$6 worst-case complexity.
• Push–Relabel: Maintain a pre-flow and a node labeling, pushing excess flow “downhill” until a feasible flow is obtained (Kannappan et al., 8 Nov 2025).

All these methods can be viewed geometrically as “circuit augmentations” in the flow space: each augmentation corresponds to a minimal-support vector in the kernel of the node-arc incidence matrix, tracing a regulated walk in the pseudoflow polyhedron (Borgwardt et al., 2023). The steepest-ascent circuit rule, for example, recovers Edmonds–Karp via maximizing the improvement per augmentation.

Recent work crystallizes this viewpoint, showing that generic path-based and push-based algorithms are unified as circuit walks with specific pivot rules (Borgwardt et al., 2023).

3. Modern Combinatorial Advances and Dense Graph Regime

A major focus of recent research is improving the complexity of exact max-flow algorithms in dense graphs ($c:E\to\mathbb{Z}_+$7). Recent combinatorial breakthroughs include a deterministic $c:E\to\mathbb{Z}_+$8 algorithm for directed max-flow with integer capacities $c:E\to\mathbb{Z}_+$9 (Bernstein et al., 20 Oct 2025), leveraging:

• Expander Hierarchies: The graph is decomposed into levels with strong expansion properties; shortcut graphs introduce additional “star” structures to facilitate global flow movement.
• Weighted Push-Relabel: A variant tailored to shortcut graphs, tracking extra combinatorial weights to guide efficient augmentation and to minimize path lengths in the augmented graph.
• Cut-Matching Games: Used for hierarchy construction, balancing expansion and the potential for sparse cuts (nonstop KRV game) (Bernstein et al., 20 Oct 2025).

These developments set new benchmarks for dense graphs, bypassing the geometric/dynamic data structures needed by continuous methods and achieving deterministic complexity for vertex capacities.

Table: Breakthrough Dense-Graph Combinatorial Algorithms

Year	Complexity	Approach
2024	$\mathbb{R}_{\geq0}$0	Expander hierarchies, weighted push-relabel
2025	$\mathbb{R}_{\geq0}$1	Simpler expander hierarchy, derandomizable

4. Structural Extensions: Regular Spaces, Higher Dimensions, and L-Bounded Flows

Regular Spaces and Generalizations

The fundamental combinatorial max-flow principles extend to regular spaces, i.e., kernels/rowspaces of totally unimodular matrices (Faigle et al., 2012). Here, “flow” generalizes to solutions in a multidimensional space subject to TU constraints. The Edmonds–Karp shortest-path principle carries over, ensuring at most $\mathbb{R}_{\geq0}$2 augmentations and preserving a submodular structure.

Simplicial Complexes

The combinatorial max-flow/min-cut duality extends from graphs to $\mathbb{R}_{\geq0}$3-dimensional simplicial complexes—a setting of chains, boundaries, and higher-homology (Maxwell et al., 2021, Latorre, 2012). In these structures:

• Flows are $\mathbb{R}_{\geq0}$4-chains with conservation via the boundary operator $\mathbb{R}_{\geq0}$5.
• Cuts correspond to certain $\mathbb{R}_{\geq0}$6-cochains annihilating prescribed cycles.
• The linear programming duality persists, but integrality and optimality split: fractional optima can arise, and computation becomes NP-hard in $\mathbb{R}_{\geq0}$7 unless the boundary matrix is totally unimodular.

L-Bounded Flows

Classical and combinatorial approaches also address path-constrained or $\mathbb{R}_{\geq0}$8-bounded flows: flows must decompose into $\mathbb{R}_{\geq0}$9-$s$0 paths of length at most $s$1 (Altmanová et al., 2019). While the problem is polynomial-time solvable as an LP, only approximate FPTAS algorithms are known combinatorially for general inputs, via exponential-length methods and path-packing duals.

5. Algorithmic Complexity: All-Pairs, Fine-Grained Theory, and Dynamic Settings

All-Pairs Max-Flow

For undirected graphs, the classic Gomory–Hu tree enables all-pairs max-flow in $s$2 time. Recent work breaks the $s$3 barrier for unit edge capacities with $s$4-time combinatorial algorithms, leveraging degree-thresholded preprocessing and partial-tree techniques (Abboud et al., 2019). For node-capacities, conditional SETH-based lower bounds are established, robust to fine-grained reductions (Trabelsi, 2023).

Fine-Grained Complexity and Non-Reducibility

The combinatorial Las Vegas algorithm for all-pairs max-flow with unit node capacities achieves $s$5 time, matching conditional lower bounds for combinatorial algorithms. Non-reducibility is enforced via fast nondeterministic verification procedures, precluding SETH-type lower bounds for this setting under NSETH (Trabelsi, 2023).

Dynamic and Parallel Max-Flow

Push–relabel-based dynamic algorithms process incremental and decremental capacity updates efficiently, particularly suited for GPU implementations. By maintaining correct excess and height labeling, small batch updates achieve significant speedups over static recomputation on real-world graphs (up to $s$6) (Kannappan et al., 8 Nov 2025).

6. Structural Characterizations and Multi-Commodity Extensions

Circuit Polyhedra and Pseudoflows

The pseudoflow polyhedron augments the feasible region of max-flow with additional slack variables, exposing minimal-support circuits corresponding to fundamental augmenting moves, both for classic and push-based schemes. This unification brings primal–dual pivoting (push=circuit, relabel=dual) under one combinatorial “umbrella” (Borgwardt et al., 2023).

Two-Commodity and Biflow Problems

For certain multicommodity flow variants, e.g., two-commodity integral biflow in undirected graphs, exact combinatorial algorithms are available in polynomial time for Seymour graphs, with solution integrality guaranteed by structural forbidden-minor characterizations. These rely on decomposition into tractable structural cases (outer-face planar, four-fan, bridge-tree, etc.) with explicit max-flow subroutines and simultaneous min-bicut identification (Ding et al., 2024).

7. Applications, Generalizations, and Open Directions

Applications

Combinatorial max-flow algorithms drive advances across domains:

• Discrete fracture network simulation: Combinatorial max-flow computes approximate fluid flow rates, efficiently handling high uncertainty and multiscale geometry (Hobé et al., 2018).
• Computer vision: Combinatorial continuous max-flow enables isotropic, null-divergence image segmentation without metrication artifacts (Couprie et al., 2010).
• Large-scale real-time analytics: Dynamic GPU algorithms enable rapid network optimization in streaming or time-evolving contexts (Kannappan et al., 8 Nov 2025).

Open Directions

Several core problems remain:

• Achieving $s$7-time combinatorial max-flow for all graph densities (Bernstein et al., 20 Oct 2025).
• Closing complexity gaps between combinatorial and continuous approaches, particularly in sparse graphs or higher-dimensional regular spaces.
• Designing exact combinatorial algorithms for L-bounded and geometric flows beyond FPTAS (Altmanová et al., 2019).
• Understanding and extending fine-grained lower bounds or non-reducibility phenomena in various capacity and commodity regimes (Trabelsi, 2023, Abboud et al., 2019).

Combinatorial max-flow persists as a canonical structure, connecting polyhedral theory, algorithm design, and the combinatorics of connectivity with broad implications in network optimization, geometry, and discrete mathematics.